3 ms·
IIRC, TPUv2 uses 16 bit floating point in some format with higher dynamic range and lower precision than standard fp16. Can someone confirm? If that is right,
by twtw 9y ago
IIRC, TPUv2 uses 16 bit floating point in some format with higher dynamic range and lower precision than standard fp16. Can someone confirm?
If that is right, is the "Tensorflow-optimized" Resnet-50 using 16bit floats when running on TPUv2?
- deepnotderp 9y agoRe: fp16 dynamic range: yes.